Domain Adaptation for Named Entity Recognition Using CRFs

Domain Adaptation for Named Entity Recognition Using CRFs Tian Tian∗,†, Marco Dinarelli∗, Isabelle Tellier∗, Pedro Dias Cardoso† ∗LaTTiCe (UMR 8094), CNRS, ENS Paris, Universit´e Sorbonne Nouvelle - Paris 3 PSL Research University, USPC (Universit´e Sorbonne Paris Cit´e) 1 Maurice Arnoux 92120 Montrouge FRANCE isabelle.tellier@univ-paris3.fr, marco.dinarelli@ens.fr † Synthesio 8-10 rue Villedo 75001 Paris FRANCE ttian, pedro@synthesio.com Abstract In this paper we explain how we created a labelled corpus in English for a Named Entity Recognition (NER) task from multi-source and multi-domain data, for an industrial partner. We explain the specificities of this corpus with examples and describe some baseline experiments. We present some results of domain adaptation on this corpus using a labelled Twitter corpus (Ritter et al., 2011). We tested a semi-supervised method from (Garcia-Fernandez et al., 2014) combined with a supervised domain adaptation approach proposed in (Raymond and Fayolle, 2010) for machine learning experiments with CRFs (Conditional Random Fields). We use the same technique to improve the NER results on the Twitter corpus (Ritter et al., 2011). Our contributions thus consist in an industrial corpus creation and NER performance improvements. Keywords: Domain Adaptation, Social Media, CRFs, Machine Learning 1. Introduction Social media (forums, Facebook, Twitter, etc) are now be- coming the major use of Internet. Within these platforms, more and more topics are being discussed everyday. The automatic analysis of these massive data is a challenge, as the texts produced in these contexts differ from previously available texts. Some works have already been dedicated to traditional Natural Language Processing (NLP) tasks such as part-of-speech tagging (Gimpel et al., 2011), and even complete parsing (Foster et al., 2011) to social media data like texts from Twitter. We focus here on the task of Named Entity Recognition (NER) with both long text and short text (like tweets). NER is a traditional NLP task (see for ex- ample the CoNLL shared task 2003 (Tjong Kim Sang and De Meulder, 2003), with texts from Reuters) which has also already been addressed for tweets (Ritter et al., 2011). In this paper, we first describe, in Section 2, a special multi- domain and multi-source NER task in English from raw data (most of them from social media) for an industrial partner. In Section 3, we explain how we created an an- notated reference corpus from four different domains and two different sources (mostly forums and Twitter) to evalu- ate the NER task. The Section 4 contains the description of our baseline, a Conditional Random Fields (CRFs) model trained with a Twitter corpus. We then conducted some do- main adaptation experiments, whose results are provided in Section 5. In these experiments, we used an iterative training method with unlabelled data from one domain. We finally suggest ways to further improve our results, espe- cially the recall value. 2. Task This work aims at helping Synthesio1 to better analyse its data automatically. Synthesio is a social listening plat- 1www.synthesio.com form, providing scalable monitoring and analytic solutions to hundreds of brands and agencies around the world. Its clients …

Content truncated to 500 words